International Journal of Applied Science and Engineering

2006. 4, 3: 275-290

A DWT Based Approach for Image Steganography

Po-Yueh Chen* and Hung-Ju Lin

Department of Computer Science and Information Engineering,

National Changhua University of Education,
No. 2 Shi-Da Road, Changhua City 500, Taiwan, R.O.C.

Abstract: In this paper we propose a new steganography technique which embeds the secret
messages in frequency domain. According to different users’ demands on the embedding capac-
ity and image quality, the proposed algorithm is divided into two modes and 5 cases. Unlike the
space domain approaches, secret messages are embedded in the high frequency coefficients re-
sulted from Discrete Wavelet Transform. Coefficients in the low frequency sub-band are pre-
served unaltered to improve the image quality. Some basic mathematical operations are per-
formed on the secret messages before embedding. These operations and a well-designed map-
ping Table keep the messages away from stealing, destroying from unintended users on the inter-
net and hence provide satisfactory security.

Keywords: Discrete Wavelet Transform; Security; Steganography.

Introduction hidden watermarks. On the other hand, steg-

anography techniques apply some cover im-
In a highly digitalized world we live today, ages to protect the confidential data from un-
computers help transforming analog data into intended internet users. According to the do-
digital forms before storing and/or processing. main where watermarks or confidential data
In the mean while, the internet develops very are embedded, both categories can be further
fast and hence becomes an important medium classified as the time domain methods and the
for digital data transmission. However, being frequency domain methods [6].
a fully open medium, the internet brought us Watermarking designs are usually consistent
not only convenience but also some hazards with the following features [7].(1) Impercep-
and risks. If the data to be transmitted are tibility: Human eyes cannot distinguish the
confidential, it is convenient as well for some difference between the watermarked image
malicious users to illegally copy, destroy, or and the original version. In other words, the
change them on the internet. As a result, in- watermarked images still preserve high image
formation security becomes an essential issue quality. (2) Security: The watermarked image
[1][2]. Various schemes for data hiding are cannot be copied, modified, or deleted by any
developed recently [3][4][5]. According to the animus observer. (3) Robustness: The water-
purposes of data hiding, these schemes are mark still can be extracted out within certain
classified into two categories: watermarking acceptable quality even the image has en-
and steganography. Watermarking is a pro- dured some signal processing or noises before
tecting technique which protects (claims) the extraction. (4) Statistically undetectable: It is
author’s property right for images by some extremely hard (or impossible) to detect the

*
Corresponding author; e-mail: [email protected] Accepted for Publication: December 10, 2006
© 2006 Chaoyang University of Technology, ISSN 1727-2394

Int. J. Appl. Sci. Eng., 2006. 4, 3 275

Po-Yueh Chen and Hung-Ju Lin

watermark by statistical and/or mathematical stego-image (the image with confidential data
analysis. (5) Blind detection: The extracting embedded in). (2) Capacity: The more data an
procedures have not to access the original image can carry the better it is. However,
image. large embedded data usually degrade the im-
For spatial domain watermarking methods age quality significantly. How one can in-
[8][9], the processing is applied on the image crease the capacity without ruining the invisi-
pixel values directly. In other words, the wa- bility is the key problem. The design of a
termark is embedded in image by modifying steganography system also can be categorized
the pixel values. The advantage of this type of into spatial domain methods [11][12] and
watermarking is easy and computationally frequency domain ones [13][14][15]. The ad-
fast. The disadvantage is its low ability to vantages and disadvantages are the same as
bear some signal processing or noises. For those we mentioned about watermarking
frequency domain methods [10], the first step methods earlier.
is to transform the image data into frequency The rest of this paper is organized as follows.
domain coefficients by some mathematical Section 2 reviews the related theoretical
tools (e.g. FFT, DCT, or DWT). Then, ac- knowledge. In section 3, the proposed steg-
cording to the different data characteristics anography method is described in details step
generated by these transforms, embed the by step. Some numerical examples are illus-
watermark into the coefficients in frequency trated as well. Experimental results and
domain. After the watermarked coefficients analysis are demonstrated in section 4. Finally,
are transformed back to spatial domain, the some concluding remarks are provided in sec-
entire embedding procedure is completed. tion 5.
The advantage of this type of watermarking is
2. Related knowledge
the high ability to face some signal processing
or noises. However, methods of this type are In this section, some preliminary techniques
computationally complex and hence slower. are reviewed including the LSB (Least Sig-
The second category of data hiding is called nificant Bits) substitution method and the
Steganography. The methods are designed to Haar-DWT. The quality of a stego-image is
embed some confidential data into some defined as well to perform the experiment
cover-media (such as texts, voices, images, analysis.
and videos). After the confidential data are
embedded, they are then transmitted together 2.1 LSB substitution
with the cover-media. The major objective is
to prevent some unintended observer from The most frequently used steganography
stealing or destroying those confidential data. method is the technique of LSB substitution
There are two things to be considered when [16]. In a gray-level image, every pixel con-
designing a steganography system: (1) Invisi- sists of 8 bits. One pixel can hence display 28
bility: Human eyes cannot distinguish the =256 variations. The weighting configuration
difference between the original image and the of an 8-bit number is illustrated in Figure 1.

2 7 26 25 2 4 23 2 2 21 2 0
Figure 1. Weighting of an 8-bit pixel

276 Int. J. Appl. Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

The basic concept of LSB substitution is to cover-image and the stego-image. The
embed the confidential data at the rightmost mathematical definition for MSE is:

) (aij  bij ) 2

bits (bits with the smallest weighting) so that
MSE  (
M N
1
M  N i 1 j 1
the embedding procedure does not affect the (4)
original pixel value greatly. The mathematical
representation for LSB method is: In Equation (4), aij means the pixel value at
xi'  xi  xi mod 2k  mi (1)
position (i, j ) in the cover-image and bij
means the pixel value at the same position in
In Equation (1), xi' represents the i th the corresponding stego-image. The calcu-
pixel value of the stego-image, xi represents lated PSNR usually adopts dB value for qual-
that of the original cover-image, and mi ity judgment. The larger PSNR is, the higher
the image quality is (which means there is
represents the decimal value of the i th block only little difference between the cover-image
in confidential data. The number of LSBs to and the stego-image). On the contrary, a small
be substituted is denoted as k. The extraction dB value of PSNR means there is great distor-
process is to copy the k-rightmost bits directly. tion between the cover-image and the
Mathematically the extracted message is rep- stego-image.
resented as:

mi  xi' mod 2k (2) 2.3 Haar-DWT

Hence, a simple permutation of the extracted The frequency domain transform we applied
mi gives us the original confidential data. in this research is Haar-DWT, the simplest
DWT [17]. A 2-dimensional Haar-DWT con-
This method is easy and straightforward.
sists of two operations: One is the horizontal
However, when the capacity is greatly in-
operation and the other is the vertical one.
creased, the image quality decreases a lot and
Detailed procedures of a 2-D Haar-DWT are
hence a suspected stego-image results. Fur-
described as follows:
thermore, the confidential data might be eas-
Step 1: At first, scan the pixels from left to
ily stolen by simply extracting the k-rightmost
right in horizontal direction. Then, perform
bits directly.
the addition and subtraction operations on
neighboring pixels. Store the sum on the left
2.2 PSNR (Peak Signal to Noise Ratio)
and the difference on the right as illustrated in
How do we determine the quality of a digital Figure 2. Repeat this operation until all the
image? Human eyes perception is the fastest rows are processed. The pixel sums represent
approach. However, although this criterion is the low frequency part (denoted as symbol L)
effective in general, the results may differ while the pixel differences represent the high
from person to person. To establish an objec- frequency part of the original image (denoted
tive criterion for digital image quality, a pa- as symbol H).
rameter named PSNR (Peak Signal to Noise
Ratio) is defined as follows:

PSNR  10 log10
2552
(3)
MSE
where MSE (Mean Square Error) stands for
the mean-squared difference between the
Figure 2. The horizontal operation on the first row

Int. J. Appl. Sci. Eng., 2006. 4, 3 277

Po-Yueh Chen and Hung-Ju Lin

Step 2: Secondly, scan the pixels from top to and we discuss in three cases based on dif-
bottom in vertical direction. Perform the addi- ferent needs of capacity (different image sizes
tion and subtraction operations on neighbor- result in different capacity as well). In fix
ing pixels and then store the sum on the top mode, there is a specific range for required
and the difference on the bottom as illustrated capacity and we discuss in two cases based on
in Figure 3. Repeat this operation until all the different image quality requirements. Let us
columns are processed. Finally we will obtain explain the working principle of these two
4 sub-bands denoted as LL, HL, LH, and HH embedding modes in the next 2 sub-sections
respectively. The LL sub-band is the low fre- respectively.
quency portion and hence looks very similar
to the original image. 3.1.1 Varying mode

Case 1： Low embedding capacity require-

ment
Suppose C is the original 8-bit gray-level
cover-image of size M C  N C . It is denoted as
C  {x ij | 1  i  M C ,1  j  N C ,
x ij  {0,1,..., 255}}
(5)
Figure 3. The vertical operation
S is the n-bit secret message represented as
The whole procedure described above is S  {si | 1  i  n, si  {0,1}} (6)
called the first-order 2-D Haar-DWT. The

n  M c  N c  8  (3/ 4)  (2 / 8) . (i.e. we exploit

first-order 2-D Haar-DWT applied on the In case 1, we assume
image “ Lena” is illustrated in Figure 4.
at most 2 bits per pixel in the 3 high fre-
quency sub-bands HH, HL, and LH). The up-
per bound is obtained by multiplying the total
number of pixels by 8 (assume 8 bits per pixel)
and 3/4 (LL sub-band is kept unaltered) and
2/8 (at most 2 LSBs are applied).
Step 1: Apply DWT on C to obtain the fre-
quency-domain matrix H. The 4 sub-bands
obtained are denoted as HLL, HHL, HLH, and
(a) (b)
HHH (All 4 sub-bands have the same size of
M C  N C / 4 ).
Figure 4. (a) Original image-Lena, (b) Result after
the first-order 2-D Haar-DWT Step 2: For the first 1  i  M C  N C bits of S,
every 2 consecutive bits are combined to form
3. Proposed method a decimal value ranging from 0 to 3. For ex-
ample, sequence 010001101111 will be
3.1. Embedding procedure transformed to sequence 101233. Every 2
consecutive values in the resulted decimal
According to different application require- sequence are further combined to perform
ments, the proposed scheme is classified as subtraction operation and form a differential
varying mode and fix mode. In varying mode, sequence ranging from -3 to +3. For example,
there is not a specific range for the capacity sequence 101233 results in a differential se-

278 Int. J. Appl.Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

quence of 1(1-0), -1(1-2), 0(3-3). As shown in more bits to distinguish the subtraction status.
Table 1, there are only 4 possible absolute In Table 1, the right Table coding is designed
values (0, 1, 2, and 3) for the elements in this to record the possible subtraction pairs. The
differential sequence. Record these absolute codes underlined are embedded in HLH and
values in HHH by substituting the 2 LSBs of the others are embedded in HHL. Embedding
coefficients in HHH with 00, 01, 10, and 11 positions in HLH and HHL are just the corre-
respectively. However, same absolute value sponding positions in HHH. The raster-scan
might be consequence of different subtraction order illustrated in Figure 5 is employed for
pairs (For example, 1 could be |3-2|, |2-1|, embedding.
|1-0|, |2-3|, |1-2|, or |0-1|). Hence, we need
Table 1. Sequence mapping table

Left table: 4 possible absolute values Right table: Status of subtraction pairs

Figure 5. Embed with raster-scan order

Figure 6 demonstrates the embedding rules

with some examples. When the value embed-
ded in HHH is 1 (there are 6 possible subtrac-
tion pairs), we have to embed 2 more bits at
the corresponding position in HLH and 1 more
bit at the corresponding position in HHL. If the
value embedded in HHH is 3 (only 2 possible
subtraction pairs), recording one more bit in
HLH is required but no bit is embedded in
HHL.
Figure 6. Values embedded in HHH determine
whether embed in HLH and HHL or not
(O denotes embedding)

Int. J. Appl. Sci. Eng., 2006. 4, 3 279

Po-Yueh Chen and Hung-Ju Lin

Step 3: The remaining bits of S: Step 5: The rounded version of E, denoted as

{M C  NC  1  i  n} are embedded at those un- F, is then stored in a specific image file for-
used LSBs (denoted as O in Figure 7) in HLH mat while K is filled in the unused tags (for
and then HHL bit by bit. For example, if the example, the ”Description” tag in TIFF for-
value embedded in HHH is 1, we cannot em- mat or the “Comment” tag in JPG format).
bed any more message bit at the correspond- The stego-image with secret message embed-
ing position of HLH but 1 more bit at the cor- ded is then ready for transmission.
responding position in HLH. The embedding
positions are illustrated in Figure 7. Case 2: Median embedding capacity re-
quirement

 M c  N c  8  (3/ 4)  (2 / 8)  n 
 
  M c  N c  8  (3/ 4)  (2 / 8) 
  M  N  8  (2 / 4)  (1/ 8) 
 c c 

Step 1~ Step 3: Identical to case 1.

Step4: Embed the remaining bits of S at the
third LSB in HLH, and HHL sub-bands. The
embedding positions are illustrated in Figure
9.
Figure 7. Remaining bits are embedded in HLH and
HHL. (O denotes available positions for
embedding and X denotes positions oc-
cupied already)

Step 4: After embedding all message bits, we

obtain the slightly modified coefficients ma-
trix H'. By performing the inverse DWT
(IDWT) on H', the stego-image E is obtained.
However, due to the LSB substitutions, some Figure 9. Remaining bits are embedded in HLH and
pixels in E are not integers ranging from 0 to HHL. (O denotes available positions for
255. As shown in Figure 8, we employ a embedding and X denotes positions
so-called “Key matrix”- K to record the 4 occupied already)
possible non-integer situations (0.0, 0.25, 0.5 Step 5: Identical to step 4 of case 1.
and 0.75). The rounded pixel values of E are Step 6: Identical to step 5 of case 1.
used to show the stego-image. In order to
perfectly reconstruct the secret message bits, Case 3: High embedding capacity requirement
K is necessary in the extracting procedure.
 M c  N c  8  [(3/ 4)  (2 / 8)  (2 / 4)  (1/ 8)]  n 
173 . 75   11 
 
  M c  Nc  8  [(3/ 4)  (2 / 8)  (3/ 4)  (1/ 8)] 
 174 . 0   00 
   
174 . 25   01 
E     K   
 174 . 5   10 
174 . 75   11 
Step 1~Step 4: Identical to case 2.
   
 175 . 0   00 
Step 5: Embed the remaining bits of S at the
third LSB in HHH sub-band. The embedding
Figure 8. Generation of the Key matrix positions are illustrated in Figure 10.

280 Int. J. Appl.Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

quence of 1(1-0), -1(1-2), 0(3-3). As shown in

Table 1, there are only 4 possible absolute
values (0, 1, 2, and 3) for the elements in this
differential sequence. Next, embed the abso-
lute values and the status values in three
sub-bands (LH, HL and HH) utilizing LSBs
substitution method. The embedding order
still uses the raster-scan order illustrated in
Figure 10. Remaining bits are embedded in HHH Figure 5 as well.
(O denotes available positions for em- In this case, we embed the absolute differ-
bedding and X denotes positions occu- ence values at two rightmost LSBs in HLH and
pied already) HHL (HLH sub-band first, then HHL sub-band)
Step 6: Identical to step 5 of case 2. and the status values at those in HHH and fur-
Step 7: Identical to step 6 of case 2. ther more (the third one, and the fourth one if
needed) LSBs in HLH and HHL sub-bands. In
3.1.2 Fix mode Table 1, the coding on the right Table is de-
signed to record the possible subtraction pairs.
The codes underlined are embedded in the
( n  M c  N c  (2 / 4)  4 ), the following two
For fix embedding capacity requirement
third or fourth LSBs in HLH and HHL
sub-bands while the others are embedded in
cases are analyzed.
HHH (and the first LSB for HLH and the second
Case 1：
for HHL if needed). The embedding positions
Step 1: Identical to step 1 of case 1 in varying in HHH are just those corresponding positions
mode.
Step 2: For the first 1  i  M C  N C bits of S,
in HLH and HHL. The embedding positions are
illustrated in Figure 11. For example, if the
every 2 consecutive bits are combined to form absolute value embedded in HLH is 3 (the first
a decimal value ranging from 0 to 3. For ex- LSB of HHH is available for embedding) and
ample, sequence 010001101111 will be that embedded in HHL is 1, then we can em-
transformed to sequence 101233. Every 2 bed the HLH status (only one bit needed) in the
consecutive values in the resulted decimal first LSB of HHH. The pairing status of HHL
sequence are further combined to perform (three status bits needed, since the value is 1)
subtraction operation and form a differential is recorded at the second LSB in HHH , the
sequence ranging from -3 to +3. For example, third, and the fourth LSBs in HHL .
sequence 101233 results in a differential se-

Figure 11. The values embedded in HLH and HHL determine whether embed in HHH or not
(O denotes embedding and X denotes no embedding)

Int. J. Appl. Sci. Eng., 2006. 4, 3 281

Po-Yueh Chen and Hung-Ju Lin

Step 3: After embedding all message bits, we sub-band) and the status values in HHH and
obtain the slightly modified coefficients ma- further more LSBs in HLH and HHL sub-bands.
trix H'. By performing the inverse DWT In Table 1, the coding on the right table is de-
(IDWT) on H', the stego-image E is obtained. signed to record the possible subtraction pairs.
Again, due to the LSB substitutions, some In this case, the codes underlined are embed-
pixels in E are not integers ranging from 0 to ded in HHH (the first and second LSBs are for
255. The “Key matrix”- K is used to record HLH status; the third and fourth ones are for
the 4 possible non-integer situations (0.0, 0.25, HHL status) the others (which are not under-
0.5 and 0.75). lined) are embedded in the third LSB in HLH
Step 4: The rounded version of E, denoted as or HHL sub-band. The embedding positions in
F, is then stored in a specific image file for- HHH are just the corresponding positions in
mat while K is filled in the unused tags (for HLH and HHL. The embedding positions are
example, the ”Description” tag in TIFF for- illustrated in Figure 12. For example, if the
mat and the “Comment” tag in JPG format). absolute value embedded in HLH is 3 and that
The stego-image with secret message embed- embedded in HHL is 1, then we can embed the
ded is then ready for transmission. HLH status (only one bit needed) in the first
LSB of HHH. The pairing status of HHL (three
Case 2： status bits needed, since the value is 1) is re-
Step 1: Identical to case 1. corded at the second LSB in HLH , the second,
Step 2: Embed the absolute values at LSBs in and the third LSBs in HHH .
HLH and HHL (HLH sub-band first, then HHL

Figure 12. The values embedded in HLH and HHL determine whether embed in HHH or not
(O denotes embedding and X denotes no embedding)
Step 3: Identical to case 1. 3.2.1 Varying mode
Step 4: Identical to case 1.
Case 1:
3.2 Extracting procedure The 8-bit gray-level stego-image of M F  NF
pixels is represented as
F  { y ij | 1  i  M F ,
The message extracting is explained as fol-

1  j  N F , y ij  {0,1,..., 255}}
lows. (7)

282 Int. J. Appl.Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

Step 1: Extract the Key Matrix from file tag Step 2: Obtain matrix H' by performing DWT
of F as transform on E (which is calculated by F +
K').
K  {k ij | 1  i  M F , Step 3: Extract the absolute values (0, 1, 2 and
1  j  N F , k ij  {00, 01,10,11}}
(8) 3) from the 2 rightmost LSBs in H'HH
sub-band. According to the value extracted,
Transform all elements of K into 0, 0.25,
LSBs of corresponding positions in H'LH and
-0.5 and -0.25 to form K', which is repre-
H'HL are used to determine the subtraction pair.
sented as
K '  {k ' ij | 1  i  M F ,
Figure 13 illustrated this rule. Base on the
mapping rules defined in Table 1, we can re-
1  j  N F , k ' ij  {0, 0.25, 0.5, 0.25}}
(9) construct 2 values (former and latter) of the
decimal sequence. Cue these decimal values
in correct order and then expand them to a
binary bit stream.

Figure 13. Extracting positions in H'LH and H'HL depend on the absolute difference value
(O denotes extraction)
Step 4: By extracting some more second maining portion of S. Cascade it with the se-
LSBs in H'LH and H'HL according to the rule quence obtained in step 3, the whole message
illustrated in Figure 14, we obtain the re- bit stream S is completely extracted.

Figure 14. Extracting positions in H'LH and H'HL

(O denotes extraction and X denotes no extraction)

Int. J. Appl. Sci. Eng., 2006. 4, 3 283

Po-Yueh Chen and Hung-Ju Lin

Case 2: Step 3: Extract the absolute values (0, 1, 2, or

Step 1~Step 4: Identical to case 1. 3) from the 2 rightmost LSBs in H'LH
Step 5: Extracting the last portion of S from sub-band. According to the value extracted,
the third LSB in H'LH and H'HL and cascade it the rightmost LSB of corresponding positions
with the result obtained in step 4. Figure 15 in H'HH and the third (together with the fourth
shows the extracting positions in this step. in some cases) LSB(s) of H'LH are used to de-
termine the subtraction pair. Figure 17 illus-
trates this rule. Base on the mapping rules de-
fined in Table 1, we can reconstruct 2 values
(former and latter) of the decimal sequence.
Cue these decimal values in correct order and
then expand them to a binary bit stream.
While H'LH sub-band is finished, keep ex-
tracting absolute values from H'HL sub-band
Figure 15. Extracting positions in H'LH and H'HL and perform the same decoding processes to
(O denotes extraction and X denotes no extraction) obtain the remaining portion of S. Cascade it
with the sequence obtained in H'LH sub-band,
Case 3: the whole message bit stream S is completely
Step 1~Step 5: Identical to case 2. extracted.
Step 6: Extracting the last portion of S from
the third LSB in H'HH and cascade it with the
result obtained in step 5. Figure 16 shows the
extracting positions in step 6.

3.2.2 Fix mode

Case 1:
Step 1: Identical to the step 1 of case 1 in
varying mode. Figure 16. Extracting positions in H'HH
Step 2: Identical to the step 2 of case 1 in (O denotes extraction and X denotes no extraction)
varying mode.

Figure 17. The extracting positions for H'LH, H'HL and H'HH
(O denotes extraction and X denotes no extraction)

284 Int. J. Appl.Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

Case 2: rules defined in Table 1, we can reconstruct 2

Step 1: Identical to case 1. values (former and latter) of the decimal se-
Step 2: Identical to case 1. quence. Cue these decimal values in correct
Step 3: Extract the absolute values (0, 1, 2, or order and then expand them to a binary bit
3) from the 2 rightmost LSBs in H'LH stream. While H'LH sub-band is finished, keep
sub-band. According to the value extracted, extracting absolute values from H'HL sub-band
the first (together with the second) LSB(s) of and perform the same decoding processes to
corresponding positions in H'HH and the third obtain the remaining portion of S. Cascade it
LSB of H'LH (if the absolute value is not 3) with the sequence obtained in H'LH sub-band,
determine the original subtraction pair. Figure the whole message bit stream S is completely
18 illustrates this rule. Base on the mapping extracted.

Figure 18. The extracting positions for H'LH, H'HL and H'HH
(O denotes extraction and X denotes no extraction)
4. Simulation results and analysis （ 258Kbyte ） ,”Boat” （ 256Kbyte ） ,”Girl”
（255Kbyte）,”Lena”（258Kbyte）,
In this section, we demonstrate the simula- and”Pepper”（256Kbyte）are employed as the
tion results and perform some data analysis. cover-images. All of them are of size
The proposed system is designed using Mat-
512 512 , 8-bit gray-level images as shown
lab 7.0 for programming. Six TIFF formatted
in Figure 19. Haar-DWT [18] is applied for
images “Airplane” （ 254Kbyte ） ,”Baboon” simplicity.

(a) Airplane (b) Baboon (c) Boat

(d) Girl (e) Lena (f) Pepper

Figure 19. Six cover-images for simulations

Int. J. Appl. Sci. Eng., 2006. 4, 3 285

Po-Yueh Chen and Hung-Ju Lin

4.1 Simulation results 4.1.1 Varying mode

The secret message S adopts pseudo-random Six resulted stego-images are listed in ap-
numbers and is denoted as: pendix. It is difficult to tell the difference
S  {si | 1  i  900 2 , si  {0,1}}
between the stego-image and the correspond-
(10) ing cover-image. Table 2 and 3 exhibit the
The performance in terms of capacity (in bit) capacity and PSNR respectively for these six
and PSNR (in dB) are demonstrated for two images in three cases.
operational modes in the following
sub-sections respectively.

Table 2. Capacity of 6 images in 3 cases

Cases Case 1 Case 2 Case 3

Images Capacity Capacity Capacity
Airplane 376710 507856 573206
Baboon 376835 507670 573392
Boat 376598 507867 573318
Girl 377038 507940 573422
Lena 376942 507856 573550
Pepper 377125 507946 573184

Table 3. PSNR of 6 images in 3 cases

Cases Case 1 Case 2 Case 3

Images PSNR PSNR PSNR
Airplane 50.8554 45.9961 44.7683
Baboon 50.7647 46.1948 44.9664
Boat 50.7499 46.1385 44.9260
Girl 50.7746 46.0763 44.8842
Lena 50.8021 46.0882 44.9011
Pepper 50.7975 46.0793 44.8973
4.1.2 Fix mode 4.2 Performance analysis

Since the capacity is fixed, only the PSNR As an example of analysis, in case 1 of
results are demonstrated here. Six resulted varying mode, we assume all embedded coef-
stego-images are listed in appendices. For ficients in HH sub-band are value 1’s which
512  512 images, The capacity is fixed at results in the worst embedding capacity. As a
256  256 (size of one sub-band) 4 (embedded result, the LH sub-band cannot embed any
bits per pixel) 2 (two sub-bands are more secret bits in the second LSB. However,
used)=524288bits for both cases. Hence, the remaining secret bits can be embedded in
PSNR (stego-image quality) is the only factor the second LSB of the whole HL sub-band.
to be analyzed. Table 4 exhibits the PSNR of Hence the first portion of embedding capacity
six images in two cases. is 256  256  4  262144 bits and the total em-
bedded capacity is 262144  65536  327680 bits.

286 Int. J. Appl.Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

This is a satisfactory amount for capacity. ones. As another example, for image “Ba-
Why the image quality becomes better while boon” in case 3 of varying mode, the worst
maintaining certain capacity value? The 4 situation is LH, HL and HH sub-bands
sub-bands of DWT coefficients in frequency changes 3 bits after the embedding of secret
domain have different properties. The pro- messages. That means the difference between
posed method keeps the coefficients in LL the original coefficients and the embedded
sub-band (which is the most important part of ones is 7. In this worst case, the PSNR is
an image) unaltered. Although we do modify 38.4694 dB. This is also a satisfactory amount
the LH sub-band (horizontal edges), HL for PSNR. To demonstrate the superiority of
sub-band (vertical edges), and HH sub-band the proposed scheme, Table 5 compares its
(diagonal edges), only the edge parts of an performance (in terms of PNSR with same
image are slightly changed. This is the main capacity) with that of the side match method
reason our approach outperforms the other [12].
Table 4. PSNR of 6 images in 2 cases

Cases Case 1 Case 2

Images PSNR PSNR
Airplane 43.2206 46.7523
Baboon 39.0033 46.5443
Boat 42.3820 46.7215
Girl 42.7842 46.8093
Lena 43.2741 46.8369
Pepper 43.4479 46.7818

Table 5. PSNR comparison with the side match method [12]

Images Lena Baboon

Methods PSNR Capacity PSNR Capacity
Three-sided 45.03dB 267242 34.93dB 483758
Proposed 52.78dB 267242 46.74dB 483758
Four-sided 48.18dB 164538 38.56dB 298413
Proposed 54.94dB 164538 52.03dB 298413

About the security issue, an untended ob- needs. Because most of coefficients in HH
server cannot correctly solve the E matrix sub-band are either 0 or very small values in
without being aware of the “Key matrix”. practical cases, some suspicion may arise
Even one may accidentally find the “Key ma- from the modification of the third LSBs.
trix” in some tags of the image format they However, the rightmost LSBs in HH are just
are just a lot of meaningless disordered num- storing the decimal values of some subtrac-
bers. Besides, the original secret messages tion operations and hence the secret messages
cannot be extracted from some rightmost are still secure.
LSBs of any sub-band. The untended observe
needs to know the mapping rules before per- 5. Conclusions and future works
forming the decoding. In case 2 and case 3 of
varying mode, embedding the third LSB in a 5.1 Conclusions
random order can provide further security

Int. J. Appl. Sci. Eng., 2006. 4, 3 287

Po-Yueh Chen and Hung-Ju Lin

A new Image Steganography scheme is tracted without the “Key matrix” and decod-
proposed in this paper. Based on different re- ing rules.
quirements of applications, two operation
modes and 5 cases are provided for selection. 5.2 Future works
According to the simulation results, the PSNR
is still a satisfactory value even the highest To reduce the extra data in the stego-images,
capacity case is applied. This is due to the We have to compress the size of ”Key ma-
different characteristics of DWT coefficients trix” as far as possible. Some novel coding
in different sub-bands. Since the most essen- schemes are available for this kind of problem.
tial portion (the low frequency part) is kept As a result, the file sizes of the original image
unchanged while the secret messages are em- and that of the corresponding stego-image
bedded in the high frequency sub-bands (cor- will not differ too much. Another issue is to
responding to the edges portion of the original efficiently integrate the proposed scheme in
image), better PSNR is not a surprising result. the JPEG2000 flow which is based on DWT
Furthermore, respectable security is main- as well.
tained as well since no message can be ex-
Appendix : the resulted stego-images
A.1. Six resulted stego-images for varying mode (case 1).

(a) Airplane (b) Baboon (c) Boat

(d) Girl (e) Lena (f) Pepper

A.2. Six resulted stego-images for varying mode (case 2).

(a) Airplane (b) Baboon (c) Boat

(d) Girl (e) Lena (f) Pepper

288 Int. J. Appl.Sci. Eng., 2006. 4, 3

 A DWT Based Approach for Image Steganography

A.3. Six resulted stego-images for varying mode (case 3).

(a) Airplane (b) Baboon (c) Boat

(d) Girl (e) Lena (f) Pepper

A.4. Six resulted stego-images for fix mode (case 1).

(a) Airplane (b) Baboon (c) Boat

(d) Girl (e) Lena (f) Pepper

A.5. Six resulted stego-images for fix mode (case 2).

(a) Airplane (b) Baboon (c) Boat

(d) Girl (e) Lena (f) Pepper

Int. J. Appl. Sci. Eng., 2006. 4, 3 289

Po-Yueh Chen and Hung-Ju Lin

References and frequency domains. Pattern Recog-

nition, 36: 969-975.
[ 1] Chang, C. C. and Chuang, L. Z. 2004. [10] Hsu, C. T. and Wu, J. L. 1999. Hidden
Image Steganography., Communication digital watermarks in images. IEEE
of the Chinese Cryptology and Informa- Transactions on Images Processing, 8:
tion Security Association (CCISA), 10, 1: 58-68.
108-122. [11] Chan, C. K. and Cheng, L. M. 2003.
[ 2] Chang, C. C. and Chuang, L. Z. 2004. Hiding data in image by simple LSB
Introduction to the Visual Cryptography, substitution. Pattern Recognition, 37:
Communication of the Chinese Cryptol- 469-474.
ogy and Information Security Associa- [12] Chang, C. C. and Tseng, H. W. 2004. A
tion (CCISA),10, 2: 1-14. Steganographic method for digital im-
[ 3] Bender, W., Gruhl, D., Morimoto, N., ages using side match. Pattern Recogni-
and Lu, A. 1996. Techniques for data tion Letters, 25: 1431-1437.
hiding. IBM Systems Journal, 35, 3 and 4: [13] Chen, T. S., Chang, C. C., and Hwang,
313-336. M. S. 1998. A virtual image cryptosys-
[ 4] Johnson, N. F. and Jajodia, S. 1998. tem based upon vector quantization.
Steganography: Seeing the Unseen. IEEE Transactions on Image Processing,
IEEE Computer, February: 26-34. 7, 10: 1485-1488.
[ 5] Lou, D. C. and Liu, J. L. 2002. Steg- [14] Chung, K. L., Shen, C. H. and Chang, L.
anography Method for Secure Commu- C. 2001. A novel SVD- and VQ-based
nications. Elsevier Science on Computers image hiding scheme. Pattern Recogni-
& Security, 21, 5: 449-460. tion Letters, 22: 1051-1058.
[ 6] Petitcolas, F. A. P. Anderson, R. J., and [15] Iwata, M., Miyake, K., and Shiozaki, A.
Kuhn, M. G. 1999. Information Hiding – 2004. Digital Steganography Utilizing
A Survey, Proceeding of IEEE, 87, 7: Features of JPEG Images, IEICE Trans-
1062-1078. fusion Fundamentals, E87-A, 4:929-936.
[ 7] Chang, C. C., Chen, T. S., and Hwang, K. [16] van Schyndel, R. G., Tirkel, A. Z., and
F. 2000. Electronic Image Techniques. Osborne, C. F. 1994. A digital water-
Taipei: Unalis. mark. IEEE International. Conf. Image
[ 8] Lu, Z. M., Wu, H. T., Xu, D. G., and Sun, Processing, 2: 86-90.
S. H. 2003. A Multipurpose Image Wa- [17] Chen, T. S., Chang, C. C., and Hwang, K.
termarking Method for Copryight Noti- F., 2002. Digital Image Processing,
fication and Protection. IEICE Transfu- Taipei: Flag.
sion. Information & Systems, E86-D, 9: [18] Chen, P. Y. and Liao, E. C. 2002. A New
September: 1931-1933. Algorithm for Haar Wavelet Transform.
[ 9] Shih F. Y. and Wu Y.T. 2003. Combina- IEEE International Symposium on Intel-
tional image watermarking in the spatial ligent Signal Processing and Communi-
cation System: 453-457.

290 Int. J. Appl.Sci. Eng., 2006. 4, 3